Discovering Unmet Needs and New Solutions with Participatory Design
The Price of Discovering Your Needs Onlineamsacta.unibo.it/5761/1/WP1116.pdfThe Price of Discovering...
Transcript of The Price of Discovering Your Needs Onlineamsacta.unibo.it/5761/1/WP1116.pdfThe Price of Discovering...
ISSN 2282-6483
The Price of Discovering
Your Needs Online
Elias Carroni
Luca Ferrari
Simone Righi
Quaderni - Working Paper DSE N°1116
The Price of Discovering Your Needs Online∗
Elias Carroni† Luca Ferrari‡ Simone Righi§¶
Abstract
Thanks to new digital technologies, web users are continuously targeted by offers
that potentially fit their interests even if they are not actively looking for a product.
Does this matching always promote transactions with high social value? We consider
a model in which web users with state-contingent preferences are targeted by relevant
banners. We characterize the optimal strategy of a seller who, in addition to the price
of the offered good, designs a banner. We show that, in equilibrium, there is a positive
relationship between the price of the offered good and the accuracy of the banner sent
to users. Then, we consider the strategic decision of a Platform that attracts sellers
because of its targeting abilities and we underline that a reduction in seller’s costs may
translate into less informative banners and lower prices,fueling purchases of goods that
rational individuals may regret due to the persuasive nature of banners.
Keywords: Bayesian Persuasion, Targeting, Platforms. JEL codes: D80, D82, D83,
L10, M37.
∗We wish to thank Ennio Bilancini, Leonardo Boncinelli, Emilio Calvano, Giacomo Calzolari, Vincenzo
Denicolo, Bruno Jullien and the participants of the seminars at the University of Firenze, Telecom Paris
Tech, DSE (Bologna), the IIBEO workshop (Alghero), and the EARIE 2017 conference (Maastricht). Si-
mone Righi acknowledges funding from the European Research Council (ERC) under the European Union’s
Horizon 2020 research and innovation programme (grant agreement No 648693).†Dipartimento di Scienze Economiche - Alma Mater Studiorum - Universita di Bologna - 1, piazza
Scaravilli, 40126 Bologna, Italy. Email: [email protected].‡University of Corsica, UMR CNRS 6240 LISA, Corte, France. Email: [email protected].§Department of Agricultural and Food Sciences - Alma Mater Studiorum - Universita di Bologna - 50,
viale Fanin, 40127 Bologna, Italy. Email: [email protected] and¶MTA TK “Lendulet” Research Center for Educational and Network Studies (RECENS), Hungarian
Academy of Sciences - H-1014 Budapest, Orszaghaz u. 30.
1
1 Introduction
Ann is surfing the web, she just started to play the guitar and suddenly, from her Facebook
feeds, new advertisements appear:1 they are related to music, music equipment and mostly
guitar lessons. Ann is a self-learner and she is captured by an advertisement of a software
which promises her quick improvements. Importantly, Ann was not actively looking for such
a product. On the contrary, because of her collected private information, the advertising
platform exposed Ann to a banner promoting a product that might fit her needs. Never-
theless, the information she acquired through the banner does not answer the question Ann
is struggling with: “do I need this product?”.
The software costs $20 and Ann knows that, if she needs the product, she would enjoy
a payoff of 30. She also knows that, if she does not need the product, she would receive a
payoff of zero and suffer a utility loss due to the fact that she would still pay for a product
that does not match her needs. Ann believes that she does not need the product with
probability 0.7. Thus, she would never buy the product at this price since her expected
payoff (0.3× (30− 20)− 0.7× 20 = -11) would be lower than zero.
What could a seller possibly do in order to convince Ann to buy the software? More
precisely, once a seller is matched with a user who is potentially interested in the sponsored
product, what are the seller’s options? One obvious option would be the one of lowering
the price of the sponsored product. For example, the seller may set a price of $9, making
Ann just indifferent between buying it or not. A different option would be to change her
perception of needs, keeping the price of $20. The seller could design an experiment which
conveys some information about Ann’s payoff-relevant state, that is, her needs. To this
purpose, the seller can commit to a signal structure over purchase recommendations.2 In
other terms, the seller designs a reliable experiment from which Ann may learn something
new about the match between the product and her needs. For instance, the seller might offer
Ann a free trial of the product, which may or may not convince her to buy it. The possible
1Realistically speaking, Ann’s browser history is partly collected in the form of cookies, text files stored
in Ann’s computer that can be accessed by other web servers that aim to use Ann’s collected information
in order to show her content which is more likely to be close to her actual needs.2Hence, the seller is a Sender with commitment power as in Rayo and Segal (2010); Kamenica and
Gentzkow (2011). We discuss the related literature in the following section.
2
recommendations to “buy” and “not buy” are sent with probability 0.95 and 0.8 conditional
on Ann’s true needs, respectively.3 Given this opportunity, Ann is willing to proceed with
the free trial. When Ann likes the experiment, i.e., she observes the recommendation “to
buy”, her expected payoff is larger than zero due to the change in her beliefs that results
from the outcome of the experiment.4
This simple example shows that, whenever an individual is uncertain about the benefits
she would derive from consuming a good, a seller might be willing to reveal some information
in order to increase the probability of selling the good. However, the example also shows
that the experiment leaves Ann with positive surplus since her expected utility is larger
than zero. This means that the seller, if allowed, could design an optimal price experiment
pair in order to maximize his profits. In other words, the price of the offered good depends
not only on how much the consumer values the product, but also on how much information
the seller is able to convey in order to reduce user’s uncertainty.
Nevertheless, the most important concern that the seller has is finding Ann. To this end,
the seller might be willing to employ the aid of a middleman who would be able to target
Ann and all users that are similar to Ann on an Internet platform. In the reality of digital
markets, advertisers outsource the job of reaching users interested in their products to
Internet giants in exchange for a fee. For instance, Facebook advertising platform explains
to its users that:
“Our ad products let businesses and organizations connect with the people who
are most likely to be interested in their products and services. We believe the
ads you see across the Facebook family of apps and services should be useful
and relevant to you.”5
Throughout this work, we focus on those situations in which the users of a website are
exposed to offers they are not actively looking for. Going back to our example, the seller
does not know Ann and the match with Ann and other similar users, who potentially share
the same interest in the product, is the result of the job of a middleman; the latter offers
3Formally: Pr(not buy|no need) = 0.8 and Pr(buy|need) = 0.95.4One can verify that whenever Ann does not like the trial, she would prefer not to buy the product.5See https://www.facebook.com/ads/about/?entry_product=ad_preferences
3
sellers the possibility to expose potentially-interested users to their offers in exchange of a
per-click fee.
This paper has a two-fold aim. The first is to characterize the optimal strategy of a
seller that consists of an information-design strategy and a price. More precisely, the seller
decides the information structure of an experiment and the price of the offered good which
can be interpreted as the price of the user’s risky action “buy the product”. We show
that the reduction of user’s uncertainty granted by the observation of the outcome of the
experiment has to be compensated by an increase of the price of the offered good. That is
to say, we derive an inverse relationship between the accuracy of the experiment and the
price of the offered good. When the experiment is fully informative the seller asks for the
maximum price he could ask for, on the contrary, when the experiment is not informative
(i.e., when it does not change user’s perception), then the seller asks for the minimum
price. In particular, we show that there exists a multiplicity of optimal experiment-price
strategies that maximize seller’s profit and always leave the user with null expected payoff.
Nevertheless, we also show that, when the production of the good as well as the design of the
experiment are costly, then the set of equilibria shrinks to a unique optimal experiment-price
pair characterized by a price that sustains a persuasive experiment. Thus, the equilibrium
experiment may suggest the user to buy a product she does not need.
Having characterized the relationship between a generic seller and a generic user, we
then switch our attention to the strategic decisions of an Internet platform maximizing
profits by selling banner spaces in exchange for a per-click fee. Sellers are lured into joining
the platform because of its targeting abilities which create potential transactions with the
users they are matched with. Accordingly, our aim is to understand the role of the platform
as a matchmaker and its effects on welfare. The first observation we make is that welfare is
maximized whenever the platform has perfect targeting abilities, as each seller is matched
with his preferred user. This allows sellers to tailor their price-experiment pair on the
basis of users’ characteristics. Users are left with null expected payoff and all surplus is
extracted. In turn, sellers’ profits are captured by the per-click fee which transfers the
rent to the platform. Since in equilibrium sellers send persuasive experiments, there may
emerge transactions generating negative social surplus. Finally, we show that a decrease
of sellers’ marginal cost of production might reduce the value created by banners. This is
4
because more efficient sellers design less informative experiments lowering the price of the
good increasing sales, also in cases in which users would have preferred not to buy. Because
a reduction of sellers’ marginal cost of production increases its profits, the platform prefers
to hire more efficient sellers whereas, in the interim, users’ decisions become cheaper and
less accurate. Therefore, the efficiency gain might be balanced by a decrease of accuracy,
thus reducing welfare.
The remainder of this paper is divided as follows. In Section 2 we present the review of
the related literature. Section 3 introduces the model and presents the optimal behavior of
a single seller (3.1) as well as the optimal choices of the platform (3.2). Section 4 is devoted
to the understanding of the impact of targeting on welfare and of the role of the platform
in information provision. Finally, Section 5 concludes.
2 Related Literature
The emergence of online markets and the so called big data have an important impact on
companies’ strategies. Indeed, the access to user-specific information by online middlemen
makes them very attractive for firms, which have the opportunity to enact tailored pricing
and advertising strategies and use these platforms to reach an enormous number of potential
customers.
Targeting is the natural way to respond to the huge mass of information accessible online.
In particular, different degrees of knowledge may lead to segment the market through price
discrimination (Thisse and Vives 1988), loyalty schemes (Shaffer and Zhang 2000), switching
offers (Chen 1997; Fudenberg and Tirole 2000; Villas-Boas 1999) and online pricing (Taylor
2004). Tailoring is thus prominent in online markets, where targeted advertising of products
may be efficient as it creates new opportunities for trade, by getting on board consumers
that would have been excluded otherwise (Bergemann and Bonatti 2011).
Digital middlemen have a key role, because of their ability to provide a two-sided match-
ing service between users and advertisers. At this regard, De Corniere and De Nijs (2016)
have demonstrated how online intermediaries may have a negative impact on welfare, as their
disclosure of consumer-specific information to advertising firms may lead to excessive prices.
5
Differently from their paper, in our model targeting is always efficient, but it generates so-
cial waste, overspending in persuasion. Moreover, many scholars have recently focused on
the sale by search engines of sponsored links to advertisers to reach online searchers. Hagiu
and Jullien (2011) show how platforms may have incentives to divert search, inducing more
search than needed. Athey and Ellison (2011) analyze and discuss the effects of different
position auctions on welfare and consumers surplus and finally, Gomes (2014) highlights a
trade-off between rent extraction and clicking volumes, as advertiser with the highest will-
ingness to pay does not necessarily offer the most relevant advertisement and long-run clicks
depend on the relevance of sponsored links. Differently from this literature, in our model
the platform is a website in which users surf to enjoy contents (e.g. newspapers, blogs) or
to interact with friends (social media) and are exposed to banners although they are not
actively looking for products. Hence, for their attention to be captured by banners, the
latter must display a relevant product. When users can be perfectly targeted with tailored
banners, the clicking volumes are always maximized and then the platform only focuses on
rent extraction. Differently, when users are not perfectly targeted, the trade-off is between
clicking volumes and the per-click fee paid by the seller to which the banner space is sold.
We borrow from the literature of targeting the matching role of online intermediaries
and their discriminating power, but we look at online targeting from another perspective.
In a context in which users have state-dependent needs, advertisement is interpreted as
a statistical experiment that is useful for users to discover their needs. Therefore, our
model interpretation of advertisement does not immediately match with the classification
traditionally adopted by the literature, i.e., persuasive (Robinson 1969; Kaldor 1950) and
informative (Stigler 1961; Nelson 1974) advertisement.6 The experiment has the only role of
providing individuals with information (more or less accurate) about their state of necessity.
In this sense, it does not entail any ex-ante change in consumer preferences but only an
ex-post demand shift due to the change in perception about the state of necessity induced
by the observation of the outcome of the experiment.
Our modeling choice is in the spirit of the recent stream of literature about information
control, initiated by Rayo and Segal (2010) and Kamenica and Gentzkow (2011). In these
6 See Bagwell (2007) for a review on different classes of advertisement.
6
models, a sender with commitment ability is able to design an experiment which reveals some
information about decision maker’s payoff relevant state. The sender designs the experiment
in order to maximize the probability with which the decision maker takes her preferred
action. Kamenica and Gentzkow (2011) study an application of their model in which an
advertiser designs an experiment in order to inform consumers about the characteristics of
the sponsored product whereas Rayo and Segal (2010) consider a case in which a platform
displays different experiments to maximize its profits exploiting their position across a
webpage. Relaxing the assumption on the commitment power of the sender, Hoffmann
et al. (2014) study a model in which the latter may decide to acquire information about
the personal characteristics of individuals and tailor messages that persuade them to take
a particular action through selective information disclosure about horizontal aspects of a
product. They find that the extent to which hyper-targeting7 may harm consumers depends
on the ability of firms to price-discriminate, on the competition between senders and on
consumers’ wariness.
3 The Model
An Internet platform runs a website and offers banner spaces to sellers in exchange for a
per-click fee. Users who visit the website are exposed to banners, which can be tailored
whenever possible. Whenever a user clicks on a banner, she receives an experiment whose
outcome can be interpreted as a purchase recommendation. The information structure
of the experiment is designed by the seller. Thus, a seller is characterized by a banner
containing the price of the good and its information. When the outcome of the experiment
is positive, the user buys the product. In line with the advertising policy of Facebook and
Google, the banners cannot contain contents that are manipulative or deceptive per se. In
our setup, this is guaranteed by the commitment power we give to the advertisers, who
cannot manipulate the outcome of the experiment.8 The timing of the model is as follows:
7Hoffmann et al. (2014) define hyper-targeting as “the collection and use of personally identifiable data
by firms to tailor selective disclosure”.8A possible source of commitment is reputation which sellers may build online thanks to the interactions
with users who frequently comment on sponsored posts. For instance, Best and Quigley (2017) show that
7
Platform
Sets per-click fees
Sellers
design experiments
Consumers
decide whether to click
All
All payoffsand set prices on the banner and possibly buy are realized
Figure 1: Timeline
First, the platform sets a per-click fee that sellers have to pay in order to advertise their
products through banners. Then, sellers enter the platform by paying the fee, and each
of them reaches targeted users. Banners contain the price of the sponsored product and
an experiment that may improve user’s knowledge of her needs. Given their prior beliefs,
consumers each consumer decides whether to click on a banner and - conditional on the
information received - whether to purchase the advertised product. Finally payoffs of all
agents are realized.
For the sake of exposition, we proceed as follows. In the next section, we study a generic
seller-user interaction. Subsequently, we study the optimal decisions of a platform that at-
tracts sellers because of its targeting abilities. We say that the platform has perfect targeting
abilities whenever it is informed about each user’s willingness to pay. When the platform
has imperfect targeting abilities, the platform is just informed about the distribution of the
willingness to pay of its users.
3.1 Offers through Banners
In this section, we formalize the example proposed in the introduction. We consider one
seller targeting a generic user with willingness to pay x ∈ [x, x]. In terms of our example, the
seller is targeting Ann and all the users who share the same characteristics x. The seller’s
product is advertised through a banner and sold at price p ≥ 0. While the user is informed
by the banner about all the product’s characteristics, she is uncertain about whether the
product will indeed be needed or not, i.e. she is characterized by an unobservable state ω ∈{0, 1}, describing her necessities. As in Johnson and Myatt (2006), the user is characterized
“the desire to persuade in the future can generate credibility, and hence persuasion, today”.
8
by state-dependent preferences u(ω, x, p). In our setup ω = 1 denotes the state of the
world in which the good is needed, while ω = 0 denotes the state of the world in which
the good is not needed. The user has a prior belief about being in the state 0, denoted as
µ ≡ Pr(ω = 0). Thus, given her prior, the user’s expected utility from buying the product
is:
UB(x, p) = −p+ (1− µ)x+ µ · 0. (1)
In the “bad” state ω = 0, the user will find the product not useful to her purposes whereas,
in the good state ω = 1, she gets the maximum payoff which is equal to her willingness to
pay minus the price. The outside option of not buying is assumed without loss of generality
to give zero payoff, UNB(x, p) = 0.
The seller reaches the user through his banner which contains the price of the offered
good and an experiment whose outcomes are statistically correlated with user’s state of
necessity. In particular, since the experiment reveals something about the user’s payoff-
relevant state, the user might find it worth clicking on the banner. Formally, an experiment
is characterized by a pair of conditional probability distributions π0 and π1:
π0 ≡ Pr(s = 0|ω = 0),
π1 ≡ Pr(s = 1|ω = 1)
where s ∈ {0, 1} denotes a shopping advice (for simplicity the message). Conditional on
the observation of message s, the user forms the posterior belief Pr(ω = 0|s) ≡ µs using
Bayes rule:
µ0 =µπ0
µπ0 + (1− µ)(1− π1),
µ1 =µ(1− π0)
µ(1− π0) + (1− µ)π1
.
In order for s to express a shopping advice of the kind “buy” or “not buy”, the experi-
ment π ≡ (π0, π1) must be incentive compatible so that the observation of s represents a
recommended action. In particular, if the message s = 1 suggests to buy the product, then
it must be the case that, conditional on message s = 1, user x weakly prefers to buy the
product, i.e. UB(x, p, π|s = 1) ≥ UNB(x, p, π|s = 1) = 0, that is:
(1− µ)π1
(1− µ)π1 + µ(1− π0)x ≥ p. (IC1)
9
Furthermore, for the experiment to be incentive compatible, conditional on s = 0,
user x needs to strictly prefer avoiding the purchase, that is UNB(x, p, π|s = 0) = 0 >
UB(x, p, π|s = 0) thus
p >(1− µ)(1− π1)
(1− µ)(1− π1) + µπ0
x. (IC0)
When the IC conditions hold, the user behavior is determined by the observation of the
result of the experiment, i.e., she is “obedient” to the latter. Notice that a user would click
on a banner only if it is incentive compatible. Otherwise she would receive recommendations
she does not agree with, and therefore she would not follow them.
Given that the seller designs the information content of the experiment but he does not
manipulate directly the observed message s, the seller is a Sender with commitment power
as in Kamenica and Gentzkow (2011). In other words, through the choice of the experiment
π ∈ [0, 1]2, the seller designs what the user learns from participating in the experiment. In
our leading example, the free trial contains some information relevant to Ann, so that she is
better informed about whether she needs the software after having tried its demo version.
The experiment can have two possible outcomes: when s = 1, it succeeds in persuading
the user that, in expected terms, she is weakly better off buying the product, whereas
when s = 0, the experiment fails to persuade her to buy it. In addition to Kamenica and
Gentzkow (2011), we allow the Sender to choose the price of the sponsored product which,
in our setup represents the cost of the risky action buy.
In order to study the relationship between user’s and seller’s decisions we proceed by
steps. First, we study the relationship between the information structure of the experiment
and the price of the offered good. We start by abstracting from the possible costs the seller
may incur in, e.g., production or shipping costs, and information-design costs. We then
relax the analysis by studying how these frictions affect seller’s optimal strategy.
Seller’s strategic variables are summarized by the banner though which he reaches the
targeted user. Thus, the seller targeting user x is described by the banner βx = (p, π) where
x is referred to the targeted user as explained above. The game develops as follows:
i) Nature draws the user’s x state of necessity ω which is unknown to both players;
ii) The seller targeting user x designs his banner βx;
10
iii) The user observes βx and decides whether to click or not on the banner;
iv) If the user does not click on the banner, she gets her reservation payoff, 0. If she
clicks, she receives a shopping advice s and then reacts accordingly.
We solve the user-seller interaction problem by means of the Bayes Correlated Equilib-
rium (BCE) (Bergemann and Morris 2016) with the Sender-Preferred refinement. Thus,
the equilibria are those BCE in which the Sender payoff is maximized, as in Kamenica and
Gentzkow (2011). In equilibrium, the banner βx = (p, π0, π1) is such that the incentive-
compatibility constraints hold. In this case the seller’s profit is given by
Π = Pr(s = 1)p. (2)
Notice that the probability with which the outcome of the experiment results in a recom-
mendation to buy the product is
Pr(s = 1) = µ(1− π0) + (1− µ)π1, (3)
which depends on the strategic choice π made by the seller. Therefore, equilibrium is an
optimal banner
β∗x = arg maxβx
Pr(s = 1)p s.t. IC0 and IC1. (4)
Proposition 1. The Optimal Banner.
Consider the seller targeting user x. The optimal banner β∗x = (p∗, π∗0, π∗1) is characterized
by π∗1 = 1 whereas π∗0 and p∗ satisfy
1− µ1− µπ∗0
x = p∗. (5)
Therefore, there is a continuum of equilibria in which π∗0(p∗) is strictly increasing in the
price of the offered good and the equilibrium payoff is (1− µ)x for any β∗x.
Proof. User’s expected payoff increases in the probability with which the experiment sends
the correct recommendation. In addition, in state ω = 1 the seller’s and the user’s interests
are perfectly aligned. Thus, because of the linearity of the problem, π∗1 = 1. Then, at π∗1 we
11
have posterior belief µ0 = 1 since message s = 0 can only be observed in state ω = 0. This
means that IC0 is always trivially satisfied. Substituting π∗1 = 1 into IC1 we are left with:
1− µ1− µπ0
x ≥ p.
The seller’s and the user’s interests are now perfectly opposed which means that, in equi-
librium, IC1 is binding. Finally, for any π0 ∈ [0, 1], there exists a p∗ ∈ R+ such that IC1 is
satisfied with equality, that is1− µ
1− µπ∗0x = p∗.
From the relationship, it is immediate to note that p∗ strictly increases in π∗0. Substi-
tuting the optimal β∗x = (p∗, π∗0, 1) into the profit equation of the seller each equilibrium
combination yields (1− µ)x.
Proposition 1 describes the choice of the optimal banner as an information-design prob-
lem in which, in addition to the information content of the experiment, the seller also
chooses the price of the product sold. At one extreme, we find the fully informative ban-
ner β∗x = (x, 1, 1) which always suggests the user not to buy when she doesn’t need the
product, and to buy in the good state of the world. As this equilibrium eliminates user’s
uncertainty about her state of necessity, the latter is willing to pay the maximal price. At
the opposite extreme, the non-informative banner β∗x = ((1− µ)x, 0, 1) suggests the buyer
to buy in all states of the world, thus effectively providing no information to the user. In
turns, the latter will be willing to buy at a price equal to her expected payoff given her
priors about the state of the world.9 Intuitively, the seller is facing an endogenous trade-off
between low demand (selling with low probability) and high prices on one side, and high
demand (selling with high probability) and lower prices on the other side. This reflects
the fact that the experiment must be incentive compatible and, therefore, the price must
accommodate the purchase recommendations. In particular, since π1 = 1, the probability
of sale resulting from a credible information structure is Pr(s = 1) = 1 − µπ0, where π0
9Observe that this would be the unique outcome of a Sender-Receiver game in which sender has no
commitment power. Given that interests are opposed, there is no information the seller could share with-
out commitment and her only possibility would be to decrease the price. Nevertheless, the commitment
assumption underlines the fact that, thanks to the platform, the seller may build some reputation which
allows her to increase the price of the good in exchange for a more accurate experiment.
12
has to satisfy the IC1. Thus, in order to increase the probability of sale, the seller has
to decrease π0, and a percentage change of π0 will give a percentage change in the prob-
ability of sale equal to ∂ Pr(s=1)∂π0
× π0Pr(s=1)
= −µπ01−µπ0 . Meanwhile, for the same percentage
change to be incentive compatible, the price has to vary and a percentage change in π0
gives ∂p∂π0× π0
p= µ(1−µ)x
(1−µπ0)2× (1−µπ0)π0
(1−µ)x= µπ0
1−µπ0 . Hence, the reduction in margins required to
provide less information (and sell also in the bad state) is always perfectly offset by the
increase in demand. As an example, consider a situation in which x = 10 and µ = 1/2. If
the seller sells only in the good state producing a fully informative experiment, π0 = 1, the
optimal price turns out to be p = 10 and the expected profit Π = 5. If the seller instead
switches to the opposite extreme, proposing a non-informative experiment with π0 = 0, the
price cannot exceed p = 5. In relation to the fully informative solution, the seller doubles
the probability of sale, but charges half the price: the expected profit would be precisely
the same as in full information.
The positive relationship between the price and π0 can be understood in terms of accu-
racy of the experiment. The accuracy of the experiment is defined as
p0
Y
µ
x
x′ > x
Figure 2: Accuracy of the experiment as a function of the price. The constraint is given by
π∗0(p) ∈ [0, 1] and µ = 0.8. Two users x′ > x receive different equilibrium banners.
13
Y := µπ0 + (1− µ)π1, Y ∈ [0, 1] (Accuracy)
that is the probability with which the purchase recommendation matches user’s needs.
Given that in equilibrium π∗1 = 1, from the relationship π∗0(p∗) ∈ [0, 1] we obtain
Y ∗ = 1− (1− µ)x
p∗− 1− µ
thus, when targeting user x, more accurate banners result in higher prices. It is worth
to stress that, keeping the price constant, the accuracy of the experiment decreases in x.
Indeed, if the user willingness to pay increases but the price remains constant, the seller
can increase the probability of sales just by designing a less accurate, though incentive
compatible, experiment.
Proposition 1 describes the abstract relationship linking the price of the offered good and
the information content of the experiment neglecting the market frictions and transaction
costs that affect seller’s behavior. Indeed, the multiplicity of optimal banners targeting
consumer x described in Proposition 1 vanishes when the seller faces a positive marginal
cost of production c > 0, which can be thought as a shipping cost from the warehouse to
the user. In particular, the seller has to pay a cost c whenever the good is purchased. Then,
the seller’s problem becomes
maxβx
Pr(s = 1)(p− c) s.t. IC0 and IC1. (6)
The cost the seller has to pay does not affect the user’s incentive compatibility constraints
which guarantee that, in equilibrium π∗1 = 1 and
1− µ1− µπ∗0
x = p∗. (7)
The next proposition shows how the set of equilibria identified in Proposition 1 shrinks in
favor of fully informative experiments whenever the seller incurs a positive marginal cost
c > 0.
Proposition 2. Marginal cost of production.
Consider the seller targeting user x. When the seller faces a marginal cost of production
c > 0, the optimal banner β∗x = (p∗, π∗0, π∗1) is characterized by π∗1 = π∗0 = 1 and p = x.
14
Proof. The relationship between p and π0 needed to fulfil incentive compatibility, i.e.,1−µ
1−µπ0x = p(π0) reduces seller’s problem to the following:
maxπ0
Pr(s = 1)[p(π0)− c] = maxπ0
[(1− µπ0)
(1− µ
1− µπ0
x− c)]
which is maximized at π∗0 = 1.
The intuition behind this result is easily grasped by following up on our previous exam-
ple. Consider again a switch from a price 10 (which is incentive compatible only if π0 = 1)
to a price of 5, but now assuming a marginal cost c = 2. The reduction in margins going
from 10 − 2 = 8 to 5 − 2 = 3 is such that the new margin is 3/8 < 1/2, whereas the
probability of sale is again doubled. This means that providing full information - and the
maximal price - is always the best solution for the seller. Therefore, a seller who wants to
target user x will do so by designing a fully informative experiment, thus selling only in the
good state at the maximal price.
Proposition 2 stresses the fact that positive costs of production provide an incentive for
the seller to design a fully informative banner in order to maximize his profits. Nevertheless,
experiments are rarely fully informative as gathering information is usually costly. Designing
an experiment would require the seller to sustain relevant costs to study the drivers that can
inform a user about her needs. In turn, those costs affect the seller’s optimal behavior in a
straightforward way as they create an additional trade-off the seller has to consider when
designing his optimal banner – the seller has less incentives to provide information given that
it is costly to do so. The cost of designing an experiment can be related to the reduction of
uncertainty the user experiences after having run the experiment (see for instance Kamenica
and Gentzkow 2014 and Martin 2017). Thus, the natural formulation of a cost function
for such a problem is based on Shannon’s entropy function (Shannon, 1948). However, in
order to not overly complicate the model, we make the following simplifying assumption
about the cost of designing an experiment: for all sellers, the cost of the experiment C(π)
is given by kC(π0) where C(·) is a strictly convex function with C(0) = 0, C ′′′(·) = 0 and
k ≥ 0 is a scale parameter which measures the ability of a seller to provide a more accurate
experiment.10 Under this formulation, all sellers are freely able to set π∗1 = 1 whereas the
accuracy of the experiment in state ω = 0 is endogenously determined. This assumption
10The fact that C ′′′(·) = 0 is a mathematical convenience that does not affect the results of the model as
15
π010
1
1− µ
Pr(ω = 1|s = 1)
C(π)
Figure 3: Posterior belief Pr(1|1) as a function of π0 and cost function. The cost of design-
ing the experiment strictly increases in π0. When π0 = 1, the experiment becomes fully
informative and the reduction of uncertainty is maximized, i.e., the user is informed about
her necessities.
16
implies that the cost of designing an experiment increases in the posterior belief 1− µ1 the
user reaches after observing message s = 1. In other words, given that π∗1 = 1, starting
from π0 = 0 one would have no reduction of uncertainty whereas ending up with π∗0 = 1
one would reach the fully informative experiment and therefore the maximum reduction of
uncertainty (see Figure 3). Thus, in this final case, the seller’s problem is
(p− c) Pr(s = 1)− kC(π0) s.t. IC0 and IC1. (8)
Proposition 3. Costly experiment.
Consider the seller targeting user x. When the seller faces marginal cost c > 0 and the cost
of designing the experiment is C(π) = kC(π0) as described above, then the optimal banner
satisfies β∗x = (p∗, π∗0, π∗1), where:
p∗ =1− µ
1− µπ0
x, π∗1 = 1, and π∗0 = C ′−1(µck
)Proof. From relationship (7), the seller’s problem reduces to
maxπ0
= (1− µ)(x− c)− µ(1− π0)c− kC(π0).
From the convexity of C(π0), it follows that the FOC
µc = kC ′(π∗0)
are necessary and sufficient and π∗0 is implicitly defined by the cost function.
When the choice of the experiment is not costly, the seller would opt for a fully informa-
tive banner whenever he faces a positive marginal cost. However, given that also designing
an experiment is costly, the seller reduces its informativeness. In addition, since the user
understands that she is going to buy the product with some probability in the bad state of
the world, the seller has to insure her by setting a lower price. This trade-off is captured
by the first order conditions, indeed:
µc︸︷︷︸MB of no sale in ω=0
= kC ′(π∗0)︸ ︷︷ ︸MC of info
. (9)
long is C is strictly convex.
17
One can think about the seller being possibly efficient along two different dimensions:
production and information design. Intuitively, a seller with a smaller marginal cost is more
efficient in producing or shipping the good, whereas a seller with a lower k is able to provide
the same level of accuracy at a smaller cost. Condition 9 identifies this relationship and the
following lemma formalizes the directions of the trade-off in terms of accuracy.
Lemma 1. Costs and Accuracy.
i) A decrease of the cost of designing the experiment results in a higher equilibrium
accuracy;
ii) A decrease of marginal costs results in a lower equilibrium accuracy;
iii) As c→ 0, the equilibrium experiment provides no additional accuracy, i.e., π∗0 → 0.
Proof. Since π∗0 = C ′−1(µck
), the equilibrium accuracy of the experiment can be written as
Y = µC ′−1(µck
)+ (1− µ)π∗1.
As C(·) is strictly convex, it directly follows that C ′−1(·) increases in k and decreases in
c. Then, since C ′(·) is monotonic, C ′−1(·) reaches its minimum, π∗0 = 0, at c = 0. Finally,
when no experiment is available, the accuracy of user’s decision is 1−µ, i.e., the probability
of buying in the good state. Therefore, the change in accuracy due to the observation of
the experiment is
Y − (1− µ) = µπ∗0
which goes to 0 when π∗0 is equal to 0.
Since the accuracy of the experiment and the price of the offered good are bounded
by the equilibrium relationship found in Proposition 1, the trade-off between the efficiency
dimensions are reflected by the price of the offered good as shown in Figure 3. A change
in µck
induces a change in the price according to the cost of designing the experiment the
seller is facing. Whereas the magnitude of the price change depends on the shape of cost
function C(·), lower marginal costs always translate into less accurate experiments, i.e., a
lower equilibrium value of π∗0.
18
π01
x
(1− µ)x
C ′(·)
p∗(π∗0)
π∗0
µck
p∗
Figure 4: The equilibrium price p∗(π∗0) is determined by the optimal choice of how much
information to convey in state 0, i.e., π0. The picture shows that the optimal level of π∗0
solves the trade-off between the cost of producing the good, c, and the marginal cost of
increasing π0, C ′(·). Then, the optimal price is determined by the equilibrium relationship
p∗ = 1−µ1−µπ∗0
x. Finally, a decrease of c, reduces the marginal cost of selling in the bad state,
π∗0 decreases and to have the offer being incentive compatible, the price has to decrease
accordingly.
19
In this section we derived the optimal banner a seller would design in order to reach a
potential consumer. The banner contains the price of the good being sold and an experiment.
If the experiment-price pair is incentive compatible, then the user clicks on the banner
and she receives a purchase recommendation behaving accordingly. We showed that the
efficiency dimensions of the seller affect the optimal equilibrium experiment-price pair. In
general, the accuracy of the seller’s experiment depends on his efficiency trade-off whereas
the price of the offered good must satisfy the incentive-compatibility constraints. Therefore,
the less accurate the experiment, the smaller the price. The analysis has nevertheless been
carried out without considering the role of the platform matching sellers and users. In the
following section, we introduce the platform who matches sellers, characterized by their
optimal banners, and users, characterized by their willingness to pay.
3.2 The platform: a perfect matchmaker
In the previous section, we characterized the optimal banner targeting user x. At this point,
we are able to discuss the role of the platform as a matchmaker. The role of the platform is
to target with a banner each user and it is precisely for this reason that sellers are willing to
pay a fee for its services. We assume that the platform sets a per-click fee f , which means
that whenever a user clicks on banner βx, the correspondent seller has to pay the amount
f . Since a banner β∗x is incentive compatible for user x, then it would generate a click also
for all users with x′ > x. Moreover, the seller providing banner β∗x is the one willing to pay
the most to reach user x in the web.11 The profit made by the seller providing banner β∗x
will be:
Π(β∗x) = (1− µ)x− (1− µπ∗0)c− C(π∗0).
In order to focus our attention on the most interesting cases, we assume the following:
Assumption 1. The marginal cost of production is sufficiently low: c > kC′(1)µ
.
11Notice that all banners βx′ with x′ < x would not generate any profit when shown to user x, who does
not click on it. Differently, all banners βx′ with x > x′ will generate a lower profit than β∗x, which is the
optimal banner that can be shown to user x, as it entirely extracts her surplus.
20
Assumption 2. x is sufficiently high: Π(β∗x) > 0.
Under assumptions 1 and 2, it is worth targeting all agents and none of them is perfectly
informed by the experiments proposed in the banners. Relaxing the first assumption would
lead to corner solutions in which the experiments are perfectly informative and, as we will
show, it will entail that the welfare gains of targeting are maximal. Relaxing the second one
would only make some people never tailored by banners, as providing them with information
would not generate any profit.
In the following, we are going to consider two different regimes: perfect and imperfect
targeting. In the first, the platform knows each user’s x and can thus perfectly tailor
banners. In the second, the platform only knows the distribution.
Imperfect targeting. The platform chooses which banner to show to each user among
the optimal banners as described in Proposition 3. The platform’s problem can be reduced
to the choice of which user x to target, since banners are optimally designed. If the platform
decides to show banner β∗x to all users, the maximal fee is f = Π(β∗x), as the fee cannot be
higher than the seller’s profit.12 Hence, the platform’s problem reduces to
maxx
[1−G(x)]︸ ︷︷ ︸click volume
Π(β∗x)︸ ︷︷ ︸per-click fee
. (10)
Intuitively, the seller faces the following trade-off. On the one hand, by targeting a user
with a lower x, the platform increases the number of clicks and reduces the fee. On the other
hand, by targeting a user with a larger x, it is able to increase the fee at the expense of a
reduction of click volume. The following lemma reports the optimal choice of the platform.
Lemma 2. When targeting is imperfect, the platform shows a unique banner β∗x, where:
x∗ = arg maxx
[1−G(x)]Π(β∗x)
Proof. To see that a solution to platform’s problem exists, notice that the function [1 −G(x)]Π(β∗x) is continuous in the closed and bounded interval [x, x], thus, by the extreme
value theorem, it attains a maximum at x∗.
12In other words, we implicitly assume that the advertising market is competitive, vis-a-vis the platform,
because sellers compete for a limited number of banner spaces. Therefore, sellers’ expected profits are zero.
The only thing that would change if sellers had market power would be the surplus distribution.
21
Perfect Targeting. In this case, the platform is able to discriminate among users choos-
ing for each x ∈ [x, x] the optimal banner β∗x which, as discussed in the previous case,
maximizes the surplus of each seller in the interaction with user x and, in turn, the rent
extractable by the platform.
Lemma 3. When targeting is perfect, each user x ∈ [x, x] is shown an optimal banner β∗x.
Proof. The proof is direct consequence of Proposition 3.
The only difference between Lemma 2 and Lemma 3 is in the number of different banners
shown by the platform. When perfect targeting is not viable, the platform has no better
option than placing the same banner βx∗ to all users. On the contrary, if individual xs are
known by the platform, each user receives a personalized offer: a banner that contains a
price and an experiment. The consequences of such a difference are better understood in
terms of welfare as we will discuss in the following section.
4 Discussion
In this section, we discuss the welfare consequences of targeting and the platform’s incen-
tives to provide accurate experiments, showing also that there is an important relationship
between accuracy of the experiments and social welfare.
In order to understand welfare, consider a user x who clicks on the banner. When the
banner gives her a positive recommendation, she buys the product receiving a payoff of
x − p in the good state and −p in the bad state. The seller receives the price and faces
a marginal cost whenever the good is sold and always pays a fee whenever the user clicks.
The platform receives the fee independently of the result of the experiment included in the
banner. As the price and the fee are transfers among agents, the total surplus generated by
each transaction user-seller is given in the good state of the world by x− c and in the bad
state by −(1− π∗0)c, as reported in the following table.
Table 1 refers to a user-seller interaction through a clicked banner. The viability of
perfect targeting affects both the clicking volumes and the per-click surplus distribution.
Targeting makes more users click on a banner. In particular, when targeting is perfect, each
22
State Prob User x Seller platform TS(x)
1 1− µ x− p p− c− f f x− c
0 µ −p(1− π∗0) (p− c)(1− π∗0)− f f −(1− π∗0)c
Table 1: Individual payoff and total surplus in each state ω from each click.
of them clicks on a personalized banner. This banner will entirely extract the tailored user’s
surplus, as the price would precisely be the one that makes her indifferent between buying
and not buying. When targeting is imperfect, the clicking volumes are lower, because users
with x < x∗ do not click on the unique banner β∗x∗ . The user with x = x∗ clicks on the banner
and receives zero expected payoff, as the tailoring banner lets the seller extract her entire
surplus. Finally, users with higher willingness to pay will receive a positive expected payoff,
as U(x > x∗, β∗x∗) > 0. Therefore, when targeting is perfect, users’ surplus is completely
extracted. Finally, the sellers are indifferent, as their surplus is always extracted by the
platform through the fully discriminating fee. The following proposition summarizes the
impact of targeting on surplus distribution.
Proposition 4. Perfect targeting entails a reduction of users’ surplus and an increase of
the platform’s profits. Sellers are always left with zero surplus.
Proposition 4 shows that perfect targeting entails a shift of consumer surplus to the
platform. This is a classic tale of two-sided price discrimination. When targeting is perfect,
each seller is matched with his desired user whose surplus in totally extracted by the banner.
In other words, perfect targeting makes all users click on a banner and possibly buy a
product, but all of them become infra-marginal. Once the sellers extract the entire surplus
of the other side of the market, the discriminating fee is used by the platform to fully take
over the value created by banners.
The reasoning used in Table 1 applies to every user-seller interaction, thus the expected
total surplus sums the surplus generated by each banner and takes into account also the
cost of designing the experiments. The expected welfare is:
23
W =
surplus︷ ︸︸ ︷(1− µ)
∫ x
x
xg(x)B(x)dx−
production cost︷ ︸︸ ︷(1− µπ∗0)c
∫ x
x
g(x)B(x)dx︸ ︷︷ ︸Wp
−information cost︷ ︸︸ ︷C(π∗0) (11)
where B(x) = 1 if agent x clicks on a banner and zero if she does not. With perfect targeting
B(x) = 1 ∀x, while with imperfect targeting B(x) = 1 for x ≥ x∗ and B(x) = 0 below
x∗. If providing any level of information were free, users would buy only in the good state,
or simply π∗1 = π∗0 = 1 and C(1) = 0, with no transactions in the bad state and maximal
number of surplus-generating transactions. Total welfare can be split into two components.
The first one is the value generated within the platform, Wp, which represents the surplus
component emerging thanks to the existence of the middleman who, through the banners,
allows the interaction between sellers and users. The second one does not concern the
platform, as it is the information cost that sellers sustain to provide potential consumers
with experiments.
Comparing the welfare with perfect and imperfect targeting yields the following propo-
sition.
Proposition 5. Welfare is strictly larger under perfect targeting. The welfare gain due to
perfect targeting drops as long as the banners become less accurate.
Proof.
When individual x’s are not known, the total welfare is given by:
W IT = (1− µ)
∫ x
x∗xg(x)dx− c(1− µπ∗0) [1−G(x∗)]− C(π∗0) (12)
With perfect targeting, the total welfare is given by:
W PT = (1− µ)
∫ x
x
xg(x)dx− c(1− µπ∗0)− C(π∗0) (13)
Comparing Equations (12) and (13) we find:
W PT −W IT = (1− µ)
∫ x∗
x
xg(x)dx− c(1− µπ∗0)G(x∗) > 0,
24
since Assumption 2 implies that the first term is larger than the second one.
Moving from imperfect to perfect targeting entails a massive boost in the clicking vol-
umes and purchases in the good as well as in the bad state. In particular, because experi-
ments are costly, fully informative experiments are never sent, i.e., π0 < 1 for all xs. Hence,
although perfect targeting increases the overall supply of information conveyed and sellers
would like to provide as much information as possible, the experiments are biased towards
persuasion, thus generating undesirable transactions in the bad state.
As a consequence, accuracy becomes very important to understand the welfare generated
by the platform by means of the banners. In particular, less accurate experiments result
in a welfare loss both under imperfect and perfect targeting. Perfect targeting emphasizes
this effect with respect to imperfect targeting, as all users, rather than some of them, click
on non-accurate banners, giving us the second part of Proposition 5.
The accuracy of experiments is then a crucial aspect as long as welfare is concerned.
Indeed, for given level of targeting, a decrease of accuracy has the effect of reducing welfare.
Therefore, two questions naturally arise. Which advertisers has the platforms incentives to
hire? Do they provide accurate experiments? Once answered these two questions, we are
able to demonstrate that increasing productive efficiency may counterintuitively decrease
the social value created by the platform, i.e., Wp in equation (11).
The platform’s objective is to maximize clicking volumes and per-click rent extraction.
Moreover, since clicking volumes essentially depend on the ability of the platform to target
users, per-click rent extraction becomes the unique objective within each targeting regime.
Therefore, the platform has always incentives to hire the seller making most profits. As
a consequence, there may exist situations in which the platform hires sellers offering less
accurate experiments, as expressed in the following proposition.
Lemma 4. Consider the platform deciding which seller to hire for the banner targeting user
x. The platform may prefer to hire sellers offering less accurate experiments.
Proof. When deciding which banner to show to user x, the platform just sets a fee that
takes the entire surplus of the corresponding seller. Since the price will always be set to
make the IC1 binding and to induce always a click, the profit of a seller is the only concern
for the platform. Since π∗1 = 1, the profit of a seller with cost c can be directly written as:
25
Π(π0|c) = (1− µ)x− (1− µπ0)c− C(π0),
where π0 is any conditional probability between zero and one. Now, consider a shock to a
lower marginal cost c′ < c. It will hold that:
Π(π0|c′) ≥ Π(π0|c) for all π0,
so that the objective function Π(π0|c′) lies above Π(π0|c) for any π0 chosen by the seller.
This also implies maxπ0
Π(π0|c′) ≥ maxπ0
Π(π0|c). Hence, the platform will always choose the
most efficient seller as it will guarantee a higher per-click fee. Combining this result with
the one in Lemma 1, hiring a more efficient seller results in less accurate experiments.
The intuition behind the result stated in Lemma 4 is that a more efficient seller can
make higher profits when producing a banner tailored for user x. The discriminating power
of the platform in setting the fees allows it to extract the entire surplus generated by
the banner. Therefore, the optimal solution for the platform is just to hire the seller
generating the highest possible surplus which is clearly reached by the most efficient seller.
An important consequence is that the platform prefers a seller proposing less accurate
experiments. Indeed, combining the results shown in Lemma 1 with the one in Lemma 4,
we show that competition for banner spaces would not result in better accuracy, since a
decrease of the sellers’ marginal cost leads, at the limit, to non-informative experiments.
On the one hand, more efficient sellers have weaker incentives to provide information about
the bad state. On the other hand, efficient sellers are more appealing to the platform, as
they achieve higher profits and consequently pay higher fees.
Lemma 4 is particularly important when we consider the effect of increasing sellers’
efficiency on the surplus that is generated within the platform. In particular, as shown
in the following proposition, a positive productivity shock may lead to a drop of surplus
brought about by banners.
Proposition 6. Assume sellers to experience a positive productivity shock, so that c de-
creases. If C ′′ < µ2ck
, this shock will result in a decrease of WP , regardless the targeting
regime.
26
Proof. Let us consider the surplus created by the platform in equation (11):
Wp = (1− µ)
∫ x
x
xg(x)B(x)dx− (1− µπ∗0)c
∫ x
x
g(x)B(x)dx. (14)
From the sellers’ first-order conditions we have that π∗0 is implicitly defined by cµ = kC ′(π∗0).
Totally differentiating the first-order condition we getdπ∗0dc
= µkC′′
. Let us consider a drop in
the marginal cost from c to c′ < c. We have:
Wp(c′)−Wp(c) = [−(1− µπ∗0(c′))c′ + (1− µπ∗0(c))c]
∫ x
x
g(x)B(x)dx.
(15)
The productivity shock does not affect clicking volumes, which will be determined by
the targeting regime. Therefore, regardless the targeting regime, Wp(c′)−Wp(c) < 0 when:
c− c′ < µ[cπ∗0(c)− c′π∗0(c′)]. (16)
Givendπ∗0dc
= µkC′′
, we have that π∗0(c) = π∗0(c′) +dπ∗0dc
(c− c′) = π∗0(c′) + µkC′′
(c− c′). Plugging
into equation (18), we get:
c− c′ < µ[c(π∗0(c′) + µ
kC′′(c− c′)
)− c′π∗0(c′)
]c− c′ < µ
[(c− c′)π∗0(c′) + µc
kC′′(c− c′)
]1 < µπ∗0(c′) + µ2c
kC′′(17)
For all π∗0(c′) ∈ [0, 1], the right-hand side is surely larger than µ2ckC′′
. Therefore, C ′′ < µ2ck
is
a sufficient condition for the positive productivity shock to lower the welfare generated by
the platform.
27
The sufficient condition found in Proposition 6 relies on the two effects triggered by a
productivity shock. The first effect is related to the efficiency gain, since costs drop, i.e.
c′ < c. The second effect emerges indirectly from Lemma 4, which expresses the positive
relationship between π∗0 and c. The two forces go in opposite directions. The direct effect
pushes Wp to increase, but the second one goes towards a welfare deterioration:
Wp(c′)−Wp(c) = (c− c′︸ ︷︷ ︸
efficiency gain
−µ[cπ∗0(c)− c′π∗0(c′)])︸ ︷︷ ︸accuracy loss
∫ x
x
g(x)B(x)dx︸ ︷︷ ︸clicking volumes
. (18)
The decrease in welfare may only result from an accuracy loss. In particular, when the
cost of designing the experiment changes quickly with π0 then a positive productivity shock
brings about a reduction of π∗0 which is small relatively to the efficiency gain of producing
at a lower cost. In Figure 4, the line C ′ becomes steeper. Conversely, as long as the cost
of designing the experiment increases slowly, the loss in terms of information overcomes
the direct efficiency gain, with detrimental consequences on the surplus originated in the
platform. In Figure 4, the line C ′ becomes less steep.
Notice that Proposition 6 is referred to the welfare that banners originate within the
platform and abstracts from the reduced cost of information. Total welfare W = Wp−C(π∗0)
must increase because of the enhanced efficiency. However, the efficiency introduced by the
productivity shock is captured by a boost in persuasion but nevertheless does not insure
an increase of socially valuable transactions. Indeed, 1− π0 is just but the probability with
which a banner “lies”. Therefore, a decrease of π0 allows the experiments to be less reliable
in exchange of lower prices.
5 Conclusion
In recent years, the evolution of Internet-market activities started to worry policy makers
insofar as digital and global transactions do not constitute an easy target for regulators.
In addition, the availability of users’ private information created a market which platforms
exploit to match sellers and users. Whereas the platform effectively increases the match
between users and sellers, its incentives do not generally imply more informed purchase
28
decisions. Regardless from its targeting ability, which determines clicking volumes, the
objective of the platform is essentially to extract a rent from sellers’ profits. The higher the
ability of sellers to produce profits, the more they become appealing to the platform.
Interestingly, whenever the increase in sellers’ profits is due to an improvement in produc-
tion efficiency, then sellers’ willingness to inform users about their needs decreases. Indeed,
an increase in sellers’ production efficiency is transferred to consumers in form of lower
prices but at the cost of more uncertain decisions. Thus, the enhancement in efficiency goes
along with a deterioration of experiments’ accuracy. As the second effect might overwhelm
the first the welfare might be reduced.
References
Athey, S. and Ellison, G. (2011). Position Auctions with Consumer Search. The Quarterly
Journal of Economics, 126(3):1213–1270.
Bagwell, K. (2007). The economic analysis of advertising. Handbook of industrial organiza-
tion, 3:1701–1844.
Bergemann, D. and Bonatti, A. (2011). Targeting in advertising markets: implications for
offline versus online media. The RAND Journal of Economics, 42(3):417–443.
Bergemann, D. and Morris, S. (2016). Bayes correlated equilibrium and the comparison of
information structures in games. Theoretical Economics, 11(2):487–522.
Best, J. and Quigley, D. (2017). Persuasion for the long run.
Chen, Y. (1997). Paying customers to switch. Journal of Economics & Management Strat-
egy, 6(4):877–897.
De Corniere, A. and De Nijs, R. (2016). Online advertising and privacy. The RAND Journal
of Economics, 47(1):48–72.
Fudenberg, D. and Tirole, J. (2000). Customer poaching and brand switching. RAND
Journal of Economics, pages 634–657.
29
Gomes, R. (2014). Optimal auction design in two-sided markets. The RAND Journal of
Economics, 45(2):248–272.
Hagiu, A. and Jullien, B. (2011). Why do intermediaries divert search? The RAND Journal
of Economics, 42(2):337–362.
Hoffmann, F., Inderst, R., and Ottaviani, M. (2014). Persuasion through selective disclosure:
Implications for marketing, campaigning, and privacy regulation. JW Goethe University
Frankfurt, UCL, and Bocconi.
Johnson, J. P. and Myatt, D. P. (2006). On the simple economics of advertising, marketing,
and product design. The American Economic Review, 96(3):756–784.
Kaldor, N. (1950). The economic aspects of advertising. The review of economic studies,
18(1):1–27.
Kamenica, E. and Gentzkow, M. (2011). Bayesian persuasion. The American Economic
Review, 101(6):2590–2615.
Kamenica, E. and Gentzkow, M. (2014). Costly persuasion. In American Economic Review
(Papers and Proceedings), volume 104, pages 457–462.
Martin, D. (2017). Strategic pricing with rational inattention to quality. Games and Eco-
nomic Behavior, 104:131–145.
Nelson, P. (1974). Advertising as information. Journal of political economy, 82(4):729–754.
Rayo, L. and Segal, I. (2010). Optimal information disclosure. Journal of political Economy,
118(5):949–987.
Robinson, J. (1969). The economics of imperfect competition. Springer.
Shaffer, G. and Zhang, Z. J. (2000). Pay to switch or pay to stay: Preference-based price
discrimination in markets with switching costs. Journal of Economics & Management
Strategy, 9(3):397–424.
30
Shannon, C. E. (1948). A mathematical theory of communication. Bell System Technical
Journal, 27(3):379–423.
Stigler, G. J. (1961). The economics of information. Journal of political economy, 69(3):213–
225.
Taylor, C. R. (2004). Consumer privacy and the market for customer information. The
RAND Journal of Economics, 35(4):631–650.
Thisse, J.-F. and Vives, X. (1988). On the strategic choice of spatial price policy. The
American Economic Review, pages 122–137.
Villas-Boas, J. M. (1999). Dynamic competition with customer recognition. The Rand
Journal of Economics, pages 604–631.
31